Parallel Implementations of Word Alignment Tool

By Qin Gao, Stephan Vogel. 2008

407 citations recorded by the source corpus.

Training word alignment models on large corpora is a very time-consuming processes. This paper describes two parallel implementations of GIZA++ that accelerate this word alignment process. One of the implementations runs on computer clusters, the other runs on multi-processor system using multi-threading technology. Results show a near-linear speed-up according to the number of CPUs used, and alignment quality is preserved.

Subfields: Artificial Intelligence, Computational Theory and Mathematics, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computational Theory and Mathematics

In Natural Language Processing

Citations in the corpus

Findings of the 2016 Conference on Machine Translation (2016, methodology): “Translation and reordering models were estimated following the Moses protocol with default setup using MGIZA++ (Gao and Vogel, 2008) for word alignment.”
The United Nations Parallel Corpus v1.0 (2016, methodology): “To speed up the word alignment procedure, we split the training corpora into four equally sized parts that are aligned with MGIZA++ (Gao and Vogel, 2008), running 5 iterations of Model 1 and the HMM model on each part.”
The Edinburgh/JHU Phrase-based Machine Translation Systems for WMT 2015 (2015, methodology): “For word alignment we used either fast_align (Dyer et al., 2013) or MGIZA ++ (Gao and Vogel, 2008), followed by the standard grow-diag-final-and symmetrization heuristic.”
FBK-UPV-UEdin participation in the WMT14 Quality Estimation shared-task (2014, methodology): “We train the word alignment models with the MGIZA++ toolkit (Gao and Vogel, 2008) implementation of the IBM models (Brown et al., 1993).”
Models and Inference for Prefix-Constrained Machine Translation (2016, methodology): “We align the bitexts with mgiza (Gao and Vogel, 2008) and estimate 5-gram language models (LMs) with KenLM (Heafield et al.”
PROMT Translation Systems for WMT 2016 Translation Tasks (2016, methodology): “We use MGIZA (Gao and Vogel, 2008) to generate word alignments.”
Meteor 1.3: Automatic Metric for Reliable Optimization and Evaluation of Machine Translation Systems (2011, methodology): “Parallel data is word aligned using the MGIZA++ toolkit (Gao and Vogel, 2008) and alignments are symmetrized using the “growdiag-final-and” heuristic.”
Word Alignment by Fine-tuning Embeddings on Parallel Corpora (2021, methodology): “• GIZA++ (Och and Ney, 2003; Gao and Vogel, 2008): an implementation of IBM models.”
The DCU-ICTCAS MT system at WMT 2014 on German-English Translation Task (2014, methodology): “In this section, alignment model is trained by MGIZA++ (Gao and Vogel, 2008) with grow-diag-final-and heuristic function.”
The TALP-UPC Machine Translation Systems for WMT19 News Translation Task: Pivoting Techniques for Low Resource MT (2019, methodology): “All other Moses conﬁguration settings are the standard ones, using KenLM as language model (Heaﬁeld, 2011; Heaﬁeld et al., 2013) and MGIZA++ (Gao and Vogel, 2008) for alignment.”
Combined spoken language translation (2014, methodology): “Word alignments for the MT track systems were created by aligning the data in both directions with MGIZA++ [26] and symmetrizing the two trained alignments.”
Chimera – Three Heads for English-to-Czech Translation (2013, positive): “A multi-threaded version exists (Gao and Vogel, 2008), which also supports incremental extensions of parallel data by applying a saved model on a new sentence pair.”
Investigations on Translation Model Adaptation Using Monolingual Data (2011, methodology): “Phrases and lexical reorderings are extracted using the default settings of the Moses toolkit.”
Is Neural Machine Translation Ready for Deployment? A Case Study on 30 Translation Directions (2016, methodology): “Their configurations were the following: The training corpora were split into four equally sized parts that were aligned with MGIZA++ [5], running 5 iterations of Model 1 and the HMM model on each part.”
Combining Domain Adaptation Approaches for Medical Text Translation (2014, methodology): “The translation and the re-ordering model utilizes the “ grow-diag-final ” symmetrized word-to-word alignments created with MGIZA++ 4 (Och and Ney, 2003; Gao and Vogel, 2008) and the training scripts from Moses.”